An assistant loses operational credibility quickly when it misreports what it can access. Pius Lanzl's system did not need a more persuasive conversational layer. It needed deterministic status checks and dependable routes for the practical work—documents, PDFs, OCR, Google tools and images.
Evidence at publication: Pius AI appeared as the primary configured agent with 12 bindings on 30 May 2026; the OpenClaw gateway, billing proxy and Ollama were active. Calendar, Drive and Sheets were recorded as connected through Composio, while Gmail and Docs were explicitly recorded as not fully active.
The company
The operating brief was a German-language business command center for Pius. Daily work could involve calendar and file access, spreadsheets, documents, incoming PDFs, scanned material and image tasks. Those are ordinary business requirements, but they cut across hosted providers and local file-processing tools.
For the system to become useful, Pius had to be able to ask naturally in German and receive an accurate account of what the assistant could do now—not a guess based on a similar connector name.
The constraint
The setup contained a mix of connected and partially connected pieces. Google Calendar, Drive and Sheets were active through Composio; Gmail and Docs were not fully active. When an assistant checked a generic search result or the wrong toolkit slug, it could report the opposite.
That made trust the limiting factor. An operator that claims a connection is missing when it is present creates unnecessary work. One that implies access where none exists creates risk. Fragile PDF behavior and repeated approval requests for basic local tasks added to the same problem.
Why existing tools were not enough
A provider connector answers only part of the question. The assistant still has to call the exact connection check, distinguish one Google toolkit from another and choose a stable local path when a file can be handled without a cloud service.
Generic model-native PDF handling was not a sufficient production path. RECAST added constrained wrappers for extraction, OCR, rendering and creation, then documented the exact status logic for Calendar, Drive, Sheets, Gmail and Docs. This reduced the space in which the agent could improvise.
Transformation map
Before
German request → generic tool search or model file handling → uncertain status → Pius verifies or retries manually
After
German request → exact toolkit or local wrapper → deterministic result → visible output → human approval only where consequential
One workflow, end to end
- Input: Pius submits a German-language PDF or scanned document request through the command surface.
- Inspection: The oc-pdf wrapper identifies the file and chooses text extraction, rendering or OCR rather than asking the model to infer the format.
- Processing: Poppler and Tesseract, including German OCR support, handle the relevant local operation.
- Context: gbrain can retain or retrieve the business context needed to interpret the material.
- Output: The assistant returns extracted content, a structured document result or a locally generated PDF through the stable wrapper path.
- Provider branch: If the task needs Google data, the system checks the exact toolkit and reports its current connection state.
- Boundary: A Gmail or Google Docs action remains unavailable until that specific connection is activated; the assistant does not present it as live.
What changed
- Built: German-first operating guidance, gbrain and active-memory configuration, PDF/OCR wrappers, document creation tools and local image routes.
- Connected: Calendar, Drive and Sheets were recorded as active through their specific Composio toolkits; the gateway, proxy and local model service were connected to the runtime.
- Tested: Toolkit-status guidance, gbrain fallback and image-model routing were checked and repaired during implementation.
- Runtime verified: The primary Pius AI agent, 12 bindings, OpenClaw gateway, billing proxy and Ollama were observed on 30 May 2026.
- Actively used: The source does not retain enough task history to publish a routine document volume or adoption rate.
- Provider or client dependency remaining: Gmail and Google Docs were not fully active. Any action through them remains dependent on the correct provider connection.
- Financial result reported by client: None retained.
Proof
- Runtime receipt: Gateway, billing proxy and Ollama were active at the verification point.
- Routing receipt: Pius AI was the primary configured agent with 12 bindings.
- Connection receipt: Calendar, Drive and Sheets were active; Gmail and Docs were explicitly identified as missing or incomplete.
- Tooling receipt: PDF information, extraction, OCR, rendering and creation routes were installed alongside German OCR support.
- Claim boundary: The record supports component and connection state, not quantified business throughput.
Technology - revealed last
The runtime uses OpenClaw 2026.5.27, Discord, billing-proxy routing, Composio MCP, gbrain, Ollama, memory-core and active-memory. Local document paths combine oc-pdf, Poppler, Tesseract with German OCR and ReportLab; creative work routes through Higgsfield and Printing Press tooling. Execution remains allowlisted so practical file work does not require unrestricted shell access.
Client quote
No source-backed verbatim client quote is published for this case. The evidence is the inspected runtime, explicit provider-state record and installed document paths.
Next transformation
See how Plutify exposed live bookkeeping data through a read-only CFO dashboard. It applies the same principle—accurate access boundaries before automation—to financial reporting.
Get your AI blueprint
If your assistant cannot reliably say which tools are connected, adding more integrations will compound distrust. Book a free AI audit to establish the first deterministic document or data route.
Source record
- OPENCLAW-DEEP-CLIENT-CASE-STUDIES.pdf, page 6.
- RECAST_CASE_STUDIES_2026_UPDATED.pdf, page 13, document RA-CASES-001, version 2.0.
- RECAST_WEBSITE_MASTER_BLUEPRINT.md, sections 10.4, 11 and 58.